UC Santa Cruz

The visualization community uses deep learning in two main ways: to explain the inner workings of deep learning models and to incorporate deep learning into visualization pipelines. In this dissertation, we focus on the latter, with a specific emphasis on feature detection within flow visualization pipelines. This dissertation underscores the efficacy of deep learning for feature detection, particularly in visualizing complex flow phenomena like rip currents and vortices. First, we explore the use of conventional flow visualization methods, such as vector clustering and timelines, to visualize rip currents. Then, we investigate the use of the appearance of the flow field for rip current detection. Subsequently, we propose a hybrid feature detection method that combines conventional flow analysis with deep learning to find rip currents by learning their behavior from short sequences of pathlines. Finally, we introduce a multimodal deep learning approach to find vortex boundaries that learn from the shape and other physical properties of pathlines or streamlines.